MCP servers

The Model Context Protocol (MCP) is an open standard for exposing tools to AI agents. Connecting an MCP server in the Library brings its tools into your workspace, where agents can call them like any other tool.

Use MCP when the capability you want already exists as an MCP server — a database connector, a search provider, an internal service — instead of writing a skill from scratch.

Transports

You can connect a server over either transport:

stdio

The runtime launches the MCP server as a local subprocess and talks to it over standard input/output. Use this for servers you run alongside Chatty (a command plus arguments and environment).

The subprocess runs inside a sandbox — the same one skill code runs in: its own namespace, a private /tmp, a read-only /usr, a fresh working directory, and an environment that holds only the variables you declared. It keeps network access, because reaching an API is what most MCP servers are for.

That sandbox needs bubblewrap (bwrap) on the host. Without it a stdio server does not start, and says so: running someone’s command unsandboxed next to the server’s own credentials is not a fallback worth having. The shipped container image includes it; if you self-host, install it or use the HTTP transport.

HTTP / SSE

The runtime connects to a remote MCP server over HTTP with Server-Sent Events. Use this for hosted servers reachable by URL.

What MCP servers expose

Once connected, a server’s tools appear in your Tools catalog and can be attached to agents. Tavily web search is integrated this way — it powers the bundled web search tool via MCP.

Adding a server

Open the MCP tab in the Library (the McpManager).

1. In the Library, open the MCP tab and choose Add server.

2. Pick the transport: stdio for a local subprocess, or HTTP/SSE for a remote server.

3. Provide the connection details:

   • stdio — the command, arguments, and any environment variables.
   • HTTP/SSE — the server URL and any required headers or auth.

4. Save and connect. The server’s tools are discovered and added to the Tools catalog.

5. Attach the new tools to an agent or wire them into a graph.

Tip

MCP tools behave like any other tool once connected — including being subject to the same agent configuration. See Tools for how tools reach agents.

In a team, registering a server takes an admin

A server’s environment and URL can name a secret, and the runtime resolves those with the workspace’s secrets. So whoever can register a server can decide where a team secret is sent. Adding, editing and deleting a team’s servers therefore takes a team admin; everyone in the team can still see them and use them.

The environment is encrypted at rest, with the same key that protects provider API keys — so it is not readable from a database dump or a backup. Rows written before that change are converted the first time somebody edits them.